Unverified50% confidenceFactExact time
The paper's threat model assumes the attacker possesses the exact same vision-language encoder and knows the geometric layout of the page.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedLarge-model-driven parsers adopt a vision-language joint understanding approach, rendering PDF pages as images, extracting layout features via a visual encoder, and then combining a language model to reason and output structured text.67% similarUnverifiedThe PAGE framework uses vision-language models to annotate each page, following a written codebook to replicate the working method of human coders66% similarUnverifiedAttackers may use geometric alignment methods to map anonymized embedding vectors back to an interpretable semantic space, potentially partially recovering the original text content.65% similarUnverifiedThe essence of prompt injection is that attackers hijack the execution logic of large language models through carefully crafted text content.64% similarUnverifiedLarge language models' capabilities in natural language understanding, code analysis, and pattern recognition make them well-suited for cybersecurity tasks such as log parsing, attack chain reconstruction, and vulnerability code review.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1014381API
curl https://kongchang.com/api/v1/knowledge/claims/1014381MCP
get_claim(id=1014381)